About Orbifold AI
Orbifold AI is building the infrastructure layer for Physical AI. As intelligent systems move beyond language into the physical world, they require a fundamentally new understanding of physics, action, and interaction.
We partner with leading robotics and world model research teams to advance the foundations of embodied intelligence, enabling intelligent systems to perceive, understand, and operate in the real world.
The standards we set, and the infrastructure we build to scale them, will define the next frontier of robotics and Physical AI.
Role Overview
This is a research seat pointed at that problem. The questions are open, they are the kind that get published, and they happen to be the questions that decide whether touch becomes a real modality for embodied learning or stays a demo.
You will also build. Answering these questions requires instrumentation that does not exist off the shelf, so you design and fabricate the sensing modules and capture rigs your own experiments depend on. The hardware is in service of the research, not the other way around, and the loop is tight because the person who fabricates the module is the person who finds its artifacts in the data.
Open Problems You Would Own
- Cross-sensor transfer. Representations and normalisation that let a model trained on one sensor work on another. Almost nothing transfers today, and this is the problem that decides whether tactile data can be a product at all.
- Verification without a reference. What does verified mean for a signal with nothing external to check it against? This is an open methodological question, and whatever answer we publish is the one partners will hold us to.
- Drift across a corpus lifetime. Sensors age, mounts shift, operators vary. Whether a tactile dataset collected over a year is internally comparable is an empirical question nobody has answered at scale, and the correction may turn out to be a modelling problem rather than a calibration one.
- Whether touch earns its bandwidth. Under what conditions the tactile channel measurably improves a manipulation policy, and where it does not. Published results are mixed and mostly on small datasets. We would be positioned to settle it.
What You Will Work On
- Run the research agenda above as real experiments: hypothesis, instrumentation, controls, and a result you would defend to a reviewer.
- Design and fabricate the instrumentation your experiments need. Compact multi-axis force sensing modules and the wearable or rig integration around them, iterated to something that survives sustained use in the field.
- Characterize rigorously. Build the test bench, measure against reference load cells, and quantify hysteresis, creep, drift, temperature sensitivity, taxel crosstalk, spatial resolution and dynamic range. Publish the numbers internally whether or not they flatter the design.
- Benchmark sensing modalities against each other across resistive, capacitive, piezoresistive, magnetic and vision-based approaches, on the manipulation tasks we care about rather than on the datasheet.
- Build calibration as a system: per-unit procedures, transfer functions, temperature compensation, and a recalibration cadence that keeps a corpus comparable over a year rather than a week.
- Solve synchronisation properly. Hardware triggering and sub-frame alignment between tactile streams sampled at kHz rates and video at frame rate, precise enough that contact onset is unambiguous.
- Define how tactile data is represented, stored, versioned and delivered alongside video, pose and action, so the channel is usable rather than merely present.
- Publish, and set the standard. Benchmarks, datasets and methods that other groups adopt are an explicit goal of this role, not a side effect of it.
What We Are Looking For
- PhD or equivalent research experience in tactile sensing, soft robotics, haptics, mechatronics or a closely related field, with first-author publications at venues such as ICRA, IROS, RSS, CoRL, RoboSoft or IEEE Transactions on Haptics.
- A published, hands-on hardware record: you have designed, fabricated and experimentally validated sensing or actuation systems, and can point at the artifacts as well as the papers.
- You design the characterization before you build the thing. Experimental rigor, honest error bars, and a habit of reporting what you measured rather than what you hoped.
- Real mechanical engineering depth. CAD, design for fabrication, materials and compliance, and a working understanding of what fails when a device is worn or gripped all day.
- Electronics fluency: PCB design, embedded firmware, sensor interfaces such as I2C and SPI, and genuine comfort at a bench with an oscilloscope, a load cell and a soldering iron.
- Strong Python and PyTorch, and enough signal processing to separate a sensor fault from a mounting fault from a software fault quickly.
We care more about the quality of the work than the number of years behind it. A recent PhD with strong first-author publications in a directly relevant area is exactly who we want to talk to.
Nice to Have
- Vision-based tactile sensing of the GelSight class, or magnetic and optical approaches.
- Visuo-tactile representation learning, or published work on cross-sensor transfer.
- Tactile simulation and contact modelling, and the sim-to-real gap specific to touch.
- An open tactile dataset, benchmark or sensor design you authored that other groups have used.
- Wearables, gloves or instrumented tooling taken to production volume, including DFM and contract manufacturing.
- Hardware-level multi-sensor synchronisation.
- Mentoring or co-supervising junior researchers.
